digitReader is aCLI Toll that classifies digits. Its features a NeUral Netowrk built from stratch with no machine learning libraries. Implementing the neural network, forward propagation, backpropagation, ReLU, softmax, cross-entropy loss, and mini-batch gradient descent without using machine-learning libraries. The trained model is packaged with the CLI and can be installed from PyPI and used with a simple command.
Touch up on the readme and installation
I’m finally on PyPi! It took way longer than I expected to turn the trained model into teh CLI tool but here we are! I also polised and edited the README. It can be installed by python3 -m pip install digitReader and run by classifyDigit –image image.jpeg. Feel free to check out my project demo!
Worked on making the woring project a comman line app that precits the digit in the given picture. Next I will write the readme
Fixed my batch training, my loss calculation was off. Also added weight saving so I dont have to retarin everytime I want to test. And of course seperated test from train! Although it still runs preetry slow 50s train runtime…. I will experinment with different batch sizes, training sizes, number of neurons, learning rates soon. I might also do a version with numpy, my original goal was to not use any libraries to learn the fundementals, but numpy is much faster than python list so I might use that to reduce runtime in a seperate version to see how much of a difference it makes!
I tried to optimize and add more trainign data but accidently made it even less acurate and take longer to process. Will have to do some research!
Trained the network with 10 images (first tried with 100 but it was SO SLOW), but it runs and makes a prediction!!!!! altough it thought 5 was a 3….. I am gonna try figure out ways to optimize my train function next fo I can train 100!
Decoding MINST images! Deconding those with no libraries are harder than I thought and I had to do a TON of research about the data files. But here it is the color value of each pixel of the first imege!! I just have to extract them all, maybe find a better way to store them so I dont have to decode them each time, and pass them to my neural network, which currently have 784 hidden input nurons, 64 hidden neurons and 10 output neurons corresponding to each digit I am trying to predict. Sorry for the super long devlog, I guess I am just excited that its turning into what I am trying to build!
I added 3 output neurons meaning the neural network will respond in tems of class 1: 0.01, class 2: 0.03, class 3: 0.96. To implement that I had to add softmax fucntion chnage my forwardPass and backwardPass implementations to work with new matrixes.
Added a second varable to the hidden neurons, and organised the code to work with it. Managed to get a very low loss!!!!!
I worked on the PCB for my Hackpad project for 2 hours and 19 minutes, but accidentally connected the Lapce recording to this project instead. I asked about the situation in #stardance-help, and support confirmed that the time cannot be reassigned due to Hackatime’s data protections. They advised me to mention the situation in both projects’ devlogs so the extra time can be accounted for during review. More details can be found here https://hackclub.slack.com/archives/C0AP0NMSP3P/p1788725484466219
Did the backtpolation to work with many x’s and y’s
Added 2 more hidden neurons, set up lists to manage their related variables and set up backpropagation for each.
I added a hidden neuron and used backpropagation to set the weights
Started practicing with a simple linear regression with 3 points!
